Integration · Framework
Kyvvu for CrewAI
A listener on your crew. Tool and model calls are checked before they run, and the crew's final answer is checked before it's returned.
Set up
A few lines, then every step is checked.
- Before every call
- CrewAI's call hooks send each tool and model call to the engine before it runs.
- Blocks abort the crew
- The blocked step is recorded and KyvvuBlockedError reaches your code.
- The final answer too
- listener.kickoff() also checks the crew's answer as an outgoing message.
from kyvvu import Kyvvufrom kyvvu_crewai import KyvvuCrewAIListenerkv = Kyvvu(api_url="https://platform.kyvvu.com", api_key="...", agent_key="crew")listener = KyvvuCrewAIListener(kv)result = listener.kickoff(crew)Two ways to start a crew
Use listener.kickoff() to withhold a blocked answer.
Tool and model calls are checked either way. Only listener.kickoff() can keep a blocked final answer from reaching the caller.
| listener.kickoff(crew) | crew.kickoff() | ||
|---|---|---|---|
| Tool and model calls checked before they run | Yes | Yes | |
| Final answer checked | Yes | Yes | |
| Blocked final answer withheld | Yes | No: CrewAI's event bus still delivers it | |
| A block reaches your code as | KyvvuBlockedError | CrewAI's own HookAborted |
When a policy says no
What a block looks like.
The docs' offline crew: a scripted model, the bundled baseline, no account. Run it with declared_tools=[] and the baseline's allowlist blocks search_web before it runs.
- Blocks abort the crew
- The blocked step is recorded, the task closes as task.error, and KyvvuBlockedError reaches your code.
- The final answer warns
- Under the bundled baseline, crew_output warns until a disclosure gate is recorded. listener.gate() from a task guardrail clears it.
- Runs as written
- A scripted model stands in for a real one, so it needs no provider key either.
"""Offline: a CrewAI crew governed by the bundled baseline, no account."""from crewai import Agent, Crew, Taskfrom crewai.llms.base_llm import BaseLLMfrom crewai.tools import toolfrom kyvvu import Kyvvu, KyvvuBlockedErrorfrom kyvvu_baseline import load_baseline_policiesfrom kyvvu.schemas import RiskClassificationfrom kyvvu_crewai import KyvvuCrewAIListenerclass ScriptedLLM(BaseLLM): """Replays scripted replies; stands in for a real model.""" replies: list[str] calls: int = 0 def call(self, messages, **kwargs): self.calls += 1 return self.replies[min(self.calls, len(self.replies)) - 1] def get_input_tokens_count(self, messages, tools=None, response_model=None): return 0 # Scripted replies use no tokens; CrewAI warns if it must estimate.@tool("search_web")def search_web(query: str) -> str: """Search the web.""" return "Article 50 of the EU AI Act sets transparency obligations."llm = ScriptedLLM(model="scripted", replies=[ 'Thought: I should search.\nAction: search_web\nAction Input: {"query": "EU AI Act Article 50"}', "Thought: I know the answer.\nFinal Answer: Article 50 sets transparency obligations.",])researcher = Agent( role="Researcher", goal="Answer questions about AI governance", backstory="You research public sources.", llm=llm, tools=[search_web],)task = Task( description="What does {topic} require?", expected_output="One sentence", agent=researcher,)crew = Crew(agents=[researcher], tasks=[task])kv = Kyvvu( agent_id="research-crew", risk_classification=RiskClassification.LIMITED, declared_tools=["search_web"], read_only_tools=["search_web"], policies=load_baseline_policies(), log_location="none", # "stdout" (the offline default) prints the trace)listener = KyvvuCrewAIListener(kv)try: result = listener.kickoff(crew, inputs={"topic": "EU AI Act Article 50"}) print(result)except KyvvuBlockedError as exc: print(f"Policy blocked: {exc}")Good to know
- Python 3.11 to 3.13, crewai 1.15.21 up to version 2. Kyvvu tests with 1.15.32.
- CrewAI 1.15.29 and later pull in mcp 1.30 or newer, which carries the fix for CVE-2026-59951. Older releases in the range pin an affected mcp.
- Declare tools by CrewAI's sanitized name: "Search Web" becomes search_web.
- Calling crew.kickoff() directly still checks the final answer, but can't withhold it: use listener.kickoff().
Next
Try it before you install it.
In your browser
The playground
Build an agent's path, pick a policy and watch the engine decide.
On your machine
Five minutes, no account
kyvvu try runs a demo agent against the bundled baseline.